Require JSON output for tools

[charleslien]

No description provided.

[charleslien]

What needs to be done:

1. Review the changes made in the "require json output for tool results" commit:

• Namely, changed ToolResultOutput from (json | media) to {value: json, media?: media[]}

2. Change all instances of ToolResultOutput[] to ToolResultOutput
3. Fix all the resulting type errors

[charleslien]

Note: perhaps the changing instances of ToolResultOutput[] to ToolResultOutput would be easier if you renamed the new one something different (like ToolResultContent)? then the typechecker would tell you where all the instances of ToolResultOutput are, since that variable no longer exists

[codebuff-team]

The core idea — moving from the old union ToolResultOutput ({type:'json',value} vs {type:'media',...}) to a single flattened shape ({value, media?}) and renaming $getToolCallString → $getNativeToolCallExampleString to reflect native JSON tool calling instead of the <codebuff_tool_call> text convention — is a reasonable direction, and the mechanical rename across all common/src/tools/params/tool/*.ts description strings is done consistently.

However this PR bundles far too much into one changeset with an empty description:

1. Unrelated agent logic rewrites: .agents/editor/best-of-n/editor-best-of-n.ts and .agents/file-explorer/file-picker.ts get substantial rewrites of extractSpawnResults/extractLastMessageText plumbing that have nothing to do with the stated title. These changes assume a new spawn-result shape (agentOutput.type === 'lastMessage') that isn't explained or justified anywhere in the diff.
2. New integration tests that call live LLM APIs (editor-best-of-n.integration.test.ts, file-explorer.integration.test.ts) require a real CODEBUFF_API_KEY, run for 60-120s, and assert loosely on substring matches in freeform model output. These are flaky and expensive to run in CI, and one of the three is already .skip'd with a comment admitting the underlying feature doesn't fully work yet.
3. cli/src/hooks/use-send-message.ts silently changes the condition for adding tool calls to subagent blocks from if (agentId) to if (parentAgentId && agentId) — a behavior change buried in a large diff with no explanation of why.
4. Cost aggregation test assertions change (73→80, 50→60 credits) without justification for why the new native tool-call format changes credit costs — this looks like tests adjusted to match new behavior rather than verifying it's correct.

Given the private repo receives hand-ported changes, this diff is too large and unfocused to port safely — the tool-output-shape change, the agent-logic rewrite, and the CLI behavior change should be three separate PRs, each with a description explaining the reasoning and edge cases handled.

[github-actions]

PR checks failed

A couple of things need fixing before this can be reviewed.
None of them are about the code itself.

Description is empty or too short.

Please say what the change does and why. If it fixes an open
issue, link it (Fixes #123). Reviewers here port accepted
changes by hand into a private source tree, so a PR that does
not explain itself is expensive to accept and usually is not.

---

Edit the PR and this check re-runs automatically.